• Non ci sono risultati.

Red and dead CANDELS: massive passive galaxies at the dawn of the Universe

N/A
N/A
Protected

Academic year: 2021

Condividi "Red and dead CANDELS: massive passive galaxies at the dawn of the Universe"

Copied!
21
0
0

Testo completo

(1)

2021-01-21T11:24:49Z

Acceptance in OA@INAF

Red and dead CANDELS: massive passive galaxies at the dawn of the Universe

Title

MERLIN, Emiliano; Fortuni, F.; Torelli, M.; SANTINI, Paola; CASTELLANO,

MARCO; et al.

Authors

10.1093/mnras/stz2615

DOI

http://hdl.handle.net/20.500.12386/29913

Handle

MONTHLY NOTICES OF THE ROYAL ASTRONOMICAL SOCIETY

Journal

490

Number

(2)

Red and dead CANDELS: massive passive galaxies at the dawn of the

Universe

E. Merlin,

1‹

F. Fortuni,

1

M. Torelli,

1

P. Santini,

1

M. Castellano,

1

A. Fontana,

1

A. Grazian,

2

L. Pentericci,

1

S. Pilo

1

and K. B. Schmidt

3

1INAF - Osservatorio Astronomico di Roma, via Frascati 33, I-00078 Monte Porzio Catone (RM), Italy 2INAF - Osservatorio Astronomico di Padova, Vicolo Osservatorio 5, I-35122 Padova, Italy

3Leibniz-Institut f¨ur Astrophysik Potsdam (AIP), An der Sternwarte 16, I-14482 Potsdam, Germany

Accepted 2019 September 11. Received 2019 August 26; in original form 2019 May 22

A B S T R A C T

We search the five CANDELS fields (COSMOS, EGS, GOODS-North, GOODS-South, and UDS) for passively evolving a.k.a. ‘red and dead’ massive galaxies in the first 2 Gyr after the big bang, integrating and updating the work on GOODS-South presented in a previous paper. We perform SED-fitting on photometric data, with top-hat star-formation histories to model an early and abrupt quenching, and using a probabilistic approach to select only robust candidates. Using libraries without (with) spectral lines emission, starting from a total of more than 20 000 z >3 sources we end up with 102 (40) candidates, including one at z= 6.7. This implies a minimal number density of 1.73± 0.17 × 10−5 (6.69± 1.08 × 10−6) Mpc−3 for 3 < z < 5; applying a correction factor to account for incompleteness yields 2.30± 0.20 × 10−5. We compare these values with those from five recent hydrodynamical cosmological simulations, finding a reasonable agreement at z < 4; tensions arise at earlier epochs. Finally, we use the star-formation histories from the best-fitting models to estimate the contribution of the high-redshift passive galaxies to the global star formation rate density during their phase of activity, finding that they account for∼5–10 per cent of the total star formation at 3 < z < 8, despite being only∼ 0.5 per cent of the total in number. The resulting picture is that early and strong star formation activity, building massive galaxies on short time-scales and followed by a quick and abrupt quenching, is a rare but crucial phenomenon in the early Universe: the evolution of the cosmos must be heavily influenced by the short but powerful activity of these pristine monsters.

Key words: Galaxies – Galaxies:evolution; Galaxies:high-redshift; Galaxies: photometry;

methods: data analysis.

1 I N T R O D U C T I O N

Quantifying the abundance of passively evolving (‘red and dead’) galaxies in the early Universe is a difficult but crucial task. We know that massive galaxies typically have red colours at all epochs: while in the local Universe this is mostly caused by the absence of young stellar populations (with a degeneracy caused by metallicity; e.g. Worthey1994), at high redshift this is more often a consequence of high star-formation rates (SFRs) coupled with strong dust obscuration (e.g. Cimatti et al.2002; Dunlop et al.2007). However, it is established that a non-negligible fraction of massive galaxies in the first∼2 Gyr after the big bang is intrinsically red because of passive evolution following the quenching of the star-formation

E-mail:emiliano.merlin@inaf.it

(SF) activity (e.g. Labb´e et al.2005; Mobasher et al.2005; Fontana et al.2009; Grazian et al.2015).

The very existence of such early red and dead massive galaxies is a challenge to our present understanding of the cosmos. The formation of the structures in the concordance -CDM cosmological scenario is inherently hierarchical (Press & Schechter1974; Lacey & Cole

1993), with large structures assembling at later times with ongoing bursts of star formation (e.g. White & Rees 1978; De Lucia & Blaizot2007). On the other hand, the so-called downsizing trend is a well-established evidence, with massive galaxies assembling their stellar content earlier, and typically on shorter time-scales, than smaller ones (Matteucci1994; Cowie et al.1996; Thomas et al.

2005; Bundy et al.2006; Cimatti, Daddi & Renzini2006). In the last decades, theoretical models and hydrodynamical simulations have struggled to reproduce the properties of the observed galactic

2019 The Author(s)

(3)

populations at all epochs (e.g. Vogelsberger et al.2014a; Feldmann et al. 2017); however, to date there is no consensus yet on a robust theoretical approach capable of accurately reconciling the observational data with the models.

The issue can also be viewed under a different perspective. In the last 15 yr the tight correlation between galaxies SFRs and stellar masses, the so-called ‘main sequence’, has become a thoroughly studied topic (Brinchmann et al.2004; Daddi et al.2007; Elbaz et al.2007; Noeske et al.2007). The sequence is now confirmed to exist to high redshifts (Rodighiero et al.2014; Schreiber et al.2015). However, at any epoch some (typically compact, bulge-dominated) galaxies fall below it, indicating little or no star formation activity and implying the occurred action of some quenching mechanism (Wuyts et al.2011; Tacchella et al.2018). The time-scales of such processes, and the physical drivers behind them, remain largely unclear to date (see e.g. Man & Belli 2018); the usual suspects include AGN-driven outflows (Brennan et al.2017), stellar feedback (Kawata1999; Chiosi & Carraro2002; Ceverino & Klypin2009; Merlin et al.2012), gas strangulation (Peng, Maiolino & Cochrane

2015), or starvation (Feldmann & Mayer 2015), virial shocking of the circumgalactic medium (Dekel & Birnboim 2006), or a combination of all these. Whatever the cause, an abrupt halt of the star formation activity makes the galaxy colours turn redder, but blue light from young stellar object can outshine the old populations for several Myr after the quenching. This makes a simple colour-based selection prone to bias, even when using rest-frame inferred magnitudes as in the UVJ diagram (Labb´e et al.2005; Wuyts et al.

2007). Galaxies that have quenched shortly before being observed will not enter the selection regions until later times unless an ad-hoc modelling is adopted, and as we have shown in our previous paper (Merlin et al.2018, M18 hereafter) in which we exploited the CANDELS photometric data for the GOODS-South field (Grogin et al. 2011; Koekemoer et al. 2011; Guo et al. 2013), this is particularly true at very high redshifts, when the time-scales of the events are comparable to the life span of the Universe. Davidzon et al. (2017) and Ichikawa & Matsuoka (2017) argue that the NUVrJ diagram is better suited to identify recently quenched galaxies. However, the CANDELS catalogue does not include a NUV band (0.23 μm), so we could not counter-check the reliability of such technique. Of course, other more refined approaches – e.g. the analysis of the main sequence where recently quenched objects are found in a transient position between star forming and passive objects – could be investigated, but the analysis would be in any case posterior to the SED-fitting, since the knowledge of the physical properties of the sources would be required.

In M18 we also showed that tailoring a reliable method to identify high-redshift passive objects is arduous anyway, because of the low signal-to-noise ratio (SNR) of such distant sources, even of the bright ones. This makes it challenging to compare the observed fluxes with template models of spectral energy distributions (SEDs) or colours. In that work, we took advantage of the photometric data from the GOODS-South catalogue, complemented with new observations (Fontana et al.2014) and new deep Spitzer mosaics; we exploited state-of-the-art techniques (T-PHOT; Merlin et al.2015,

2016) and ad-hoc SED-fitting libraries built with constant (a.k.a. ‘top-hat’) star formation histories (SFHs); and we used a stringent statistical approach to exclude potential false positives. In this way we ended up with 30 passive candidates at z > 3. However, we also showed how changing some properties of the stellar libraries, or letting the redshifts of the solutions vary, dramatically impacted the results, reducing the sample to 10 (including nebular line emission in the models) or even only two (letting the redshifts free in the fitting

process) candidates. To strengthen the robustness of our selection and validate the basic assumptions, in Santini et al. (2019, S19 hereafter) we further checked the nature of the M18 candidates by means of archival ALMA data (available for 26 out of the 30 sources), statistically corroborating the passive classification of the sample from the lack of on-going star formation as seen at sub-mm wavelengths, free from the parameter degeneracies (especially the age-dust degeneracy) typical of the optical domain. Moreover, we could robustly confirm the individual passive nature of 35 per cent of our candidates, adopting conservative assumptions. In M18 we also showed that upcoming facilities such as the

James Webb Space Telescope will propel a leap forward, allowing

for a much more robust photometric precision and, consequently, determination of physical properties. However, for now we can only trust the predictive and analytic power of the currently available instrumentation, and enlarge the statistical significance including more data.

To this aim, in this paper we discuss the results from the joint analysis of the remaining four CANDELS fields (COSMOS, EGS, GOODS-North, and UDS). Since we used a refined grid of SED models, we also repeat the processing on GOODS-South. The five fields have different typical depths, therefore mixing the analysis might be risky, but we can safely consider the results as a lower limit to the actual number of passive objects above z > 3. Furthermore, in this work we address two more points: the concordance of the observations with the predictions from numerical models, and the impact that these early monsters had on the global SFH of the Universe.

The paper is organized as follows. In Section 2 we describe the data set and we briefly summarize the method we used to single out the passive sample. In Section 3 we discuss the confirmation of the candidates by means of the available far-infrared (FIR) and spectroscopic data, and in Section 4 we discuss some properties of our candidate galaxies. In Section 5 we compute the number densities of our passive sample, and we compare our findings to the predictions of five state-of-the-art hydrodynamical models:

ILLUSTRIS (Vogelsberger et al. 2014a), ILLUSTRIS-TNG100 and TNG300 (Pillepich et al.2018),EAGLE(Schaye et al.2015), and

SIMBA(Dav´e et al. 2019). In Section 6 we present a method to compute the star formation rate density (SFRD) from the fitted SEDs, and compare the contribution of the red and dead populations to the total. Finally, in Section 7 we summarize and discuss the main findings of the work.

Throughout the paper, we assume a -CDM cosmology (H0=

70.0, = 0.7, m= 0.3), a Salpeter (1959) initial mass function

(IMF) except where noted otherwise, and AB magnitudes.

2 DATA S E T A N D M E T H O D S

For the GOODS-South field we use again the 19 bands catalogue already discussed in M18, which improves on the original catalogue published by Guo et al. (2013) as it includes three more bands (WFC3 F140W from the Hubble Space Telescope and VIMOS B, plus the deep HAWK-I Ks band presented in Fontana et al.2014), and it has improved photometry on the Spitzer bands thanks to new mosaics (IRAC CH1 and CH2, by R. McLure) and new software (all four channels were re-processed usingT-PHOT). As anticipated we decided to re-analyse the GOODS-South field taking advantage of refined SED libraries and redshift estimates (see below).

For the remaining four fields, we exploited the published CAN-DELS photometric catalogues, released in 2015 and presented in Nayyeri et al. (2017), Stefanon et al. (2017), Barro et al. (2019),

(4)

Table 1. Summary of the five CANDELS catalogues: number of

photomet-ric bands, 5σ limiting magnitudes in WFC3 F160W (the detection H band), area in arcmin2. In COSMOS, 20 out of 43 bands are medium/narrow bands

from SuprimeCam and NEWFIRM; in EGS, 6 out of 23 bands are medium bands from NEWFIRM. In GOODS-North the estimate of the limiting magnitude corresponds to the Wide area; corresponding limiting magnitudes for the intermediate and deep areas are 28.2 and 28.7, respectively. A similar pattern applies to GOODS-South, where the estimate of the limiting magnitude corresponds to the Wide area; corresponding limiting magnitudes for the Deep and HUDF areas are 28.16 and 29.74, respectively. mlimare

typically defined as 5× the average standard deviation within 2 FWHM circular apertures in empty regions of the fields.

Field Bands Hlim Area

COSMOS 43 (20) 27.56 216.0

EGS 23 (6) 27.6 206.0

GOODS-N 18 27.8 173.0

GOODS-S 19 27.36 173.0

UDS 19 27.45 201.7

and Galametz et al. (2013) for COSMOS, EGS, GOODS-North, and UDS, respectively. All catalogues are based on ∼20 wide bands, and in COSMOS and EGS they are complemented with some narrow and/or medium bands. Fluxes have been typically measured by means of SEXTRACTOR (Bertin & Arnouts 1996) aperture photometry for Hubble bands, after PSF-matching to the detection band H160; and template-fitting for ground-based and

Spitzer bands, with TFIT (Laidler et al.2007) orT-PHOT(Merlin et al.2015,2016).1The typical depth of the detection band, WFC3

F160W, is∼27.5 (5σ in 2 FWHM or full width at half-maximum

diameter). The properties of the five fields are summarized in Table1; the cumulative area is∼969.7 arcmin2.

As for the redshifts, we took advantage of the latest CANDELS estimates, to be presented in Kodra et al. (in preparation) which improve upon the original Dahlen et al. (2013) estimates; the new photo-z’s (where spec-z’s are not available) are obtained combining four independent estimations, using the minimum Frechet distance combination method.

In COSMOS, ID-16676 corresponds to the Z-FOURGE source 20115, a recognized passive source first discussed in Glazebrook et al. (2017), and object of a thorough study by Schreiber et al. (2018a), who showed how the presence of sub-mm flux is actually due to a strongly obscured close companion. Glazebrook et al. (2017) assigns to this source a spectroscopic redshift of 3.7172, different from both the CANDELS and the 3D-HST photo-z’s (4.127 and 3.545, respectively). We take their spec-z as the reference redshift of the object and use it in all our subsequent analysis. In other cases of spectroscopically confirmed redshift we kept the CANDELS photo-z since the estimate was always sufficiently close (see Section 3.2).

1In template-fitting techniques, cutouts from the high-resolution detection

band are used as priors to build low-resolution templates of the sources, by means of a convolution kernel that matches the PSFs of the two images. The templates are then used to solve a linear system minimizing the difference between a model collage and the real low-resolution image, assigning to each source a multiplicative factor that best matches the observed flux. The method has proven to yield great improvements especially when the blending of the sources becomes important, as it is the case for ground-based and mid/far-infrared bands.T-PHOTis the heir of TFIT; it improves on it in terms of accuracy, robustness, and computational performance, and it includes a number of additional options. For a detailed description of the techniques and the codes, please refer to the cited publications.

2.1 Selection technique and results

As we did in M18, we proceed as follows to single out our red and dead candidates:

(i) we build a library of SED templates with top-hat SFHs, to better model the abrupt quenching of star formation in the very early Universe. The rationale for this choice is extensively discussed in M18. We improved upon the library used in the previous work by extensively refining the grid of models; the new library is described in detail in Appendix A;

(ii) we consider all the objects from the H-detected catalogues of the five CANDELS fields at zCANDELS≥ 3.0, and we select the

objects with H < 27 and with SNR >1 in Ks, IRAC-CH1, and IRAC-CH2;

(iii) on these lists of sources, we perform SED-fitting using our codeZPHOT(Fontana et al.2000), in two flavours: (i) without the inclusion of nebular emission lines (we dub resulting selection ‘reference’ sample), and (ii) with the inclusion of the lines (we dub this selection ‘lines’ sample). We also made a third run, (iii) including the lines and letting the redshift of the fit vary (‘z-free’ sample), as in M18. However, S19 have shown that the third criterion is too conservative, given that in M18 only 2 over 30 galaxies survived this selection for the GOODS-South field, while 9 out of 26 sources have been robustly confirmed as passive by means of the follow-up analysis on ALMA data. Therefore, we only cite it here for the sake of completeness, but we will not discuss it further in the paper;

(iv) we include in our lists of passive candidates only the objects having: (i) a passive best-fitting model with probability pbest>

30 per cent; and (ii) only star-forming solutions (if they have any) with probability pSF<5 per cent (these figures come from tailored

simulations which are discussed in M18).

In this way we end up with the three selections summarized in Table2. In total, we find 102 candidates in the ‘reference’ sample, which become 40 in the ‘lines’ selection. In Fig. 1we show the positions of the ‘reference’ candidates on the five CANDELS fields. We point out that while the properties of the detection band (H160) are quite similar in the five fields, the same does not hold for what concerns the other bands. In Fig.2the nominal 5σ limiting magnitudes across the spectrum, as reported in the papers describing the catalogues, are compared: they show that the quality is far from uniform. This is particularly true for the K and Spitzer bands, which are crucial for the characterization of high-z objects:

(i) the K bands come from various surveys and facilities, so that their properties are very different in the five fields. For example, the HAWK-I Ks mosaic in GOODS-South is very deep (mlim  27.0,

5σ in 2 FWHM) with the finest seeing (FWHM  0.4 arcsec), ensuring exquisite quality data. On the other hand, UDS HAWK-I

Ks has FWHM 0.4 arcsec and mlim 25.9, the COSMOS VISTA

Ks band has FWHM 0.98 arcsec and mlim 24.8, EGS WIRCAM

has FWHM 0.65 arcsec and mlim  24.3, and GOODS-North

CHFT WIRCam has FWHM 0.6 arcsec and mlim  24.7. The

filter response functions are different as well;

(ii) the IRAC bands also reach very different depths in the various fields: for example, CH1 reaches mlim= 26.5 at 5σ in GOODS-S,

while it is limited to 24.7 in UDS, 24.5 in GOODS-N, 24.4 in COSMOS, and 23.9 in EGS.

Of course, these differences have a strong impact on the efficiency of our methods in the five fields, resulting in significant variations of the number of candidates, as Table2shows.

(5)

Table 2. Number of sources in the five CANDELS fields, and in total. Left to right, we list: the

total number of detected sources, the number of sources at zCANDELS>3, the number of sources

at zCANDELS>3 with SNR>1 in K, IRAC1, and IRAC2; and the number of selected candidates in

the two passive selections: the ‘reference’ sample and the ‘lines’ sample obtained including nebular lines in the library.

Field/Sample Total z >3 S/Nz >3>1 Reference Lines

COSMOS 38 671 3778 1525 4 2 EGS 41 457 4830 1775 13 5 GOODS-N 35 445 3953 1793 36 11 GOODS-S 34 930 5029 2884 33 13 UDS 35 932 4018 2540 16 9 All fields 186 435 21 608 10 517 102 40

It is worth pointing out that the new redshift estimates change the selection for the GOODS-South field with respect to our sample in M18. Of the present 33 candidates, 26 are in common with the previous selection; 4 are now excluded (IDs 5592, 9091, 26802, 10759) and 7 new are included (IDs 3718, 4202, 4949, 5934, 13394, 16526, 19883). All of the changes are due to variations in the probabilities of star forming solutions in the SED-fitting procedure caused by the different photo-z. Since we checked that our candidates were reliable both in M18 and in S19, this seem to imply that our selection criteria are conservative, and the true number of passive objects is probably higher than these estimates.

We also note that IDs 2075 in COSMOS; 24177 in EGS; 157, 643, 6620, 9626, 13007, 24572, and 27251 in GOODS-North; 4949, 6407, 12178, and 19446 in GOODS-South, show complex morphologies in the optical, H or K bands, and possibly have close companions which might cause strong contamination despite the robustness of the adopted photometric methods.

3 C O N F I R M AT I O N O F C A N D I DAT E S W I T H FA R - I N F R A R E D A N D S P E C T R O S C O P I C DATA In this Section we study the properties of our selected candidates using their far-infrared photometry and spectroscopic data, when available.

3.1 Herschel fluxes

As in M18, we have cross-matched the positions of our selected can-didates with Herschel public catalogues, to check for degeneracies and possible misinterpretations of low-z dusty galaxies as high-z passive dust-free sources. FIR data are available for all CANDELS fields.

For the two GOODS fields we take advantage of the new, deep

ASTRODEEP catalogues by Wang et al. (in preparation), which combine data from PEP (Lutz et al.2011) and GOODS-Herschel (Elbaz et al.2011) surveys with the PACS camera, and the HerMES survey (Oliver et al.2012) with SPIRE. The catalogues are extracted using H-band CANDELS priors, making the cross-correlation very handy. We find that all the new candidates in GOODS-S have no match, while we refer the reader to M18 for the discussion on the two candidates having a FIR-counterpart (IDs 3973 and 10578). The fact that these two sources, also showing X-ray emission, are among the ones robustly confirmed by ALMA (see S19), consolidates their interpretation as galaxies that have been recently quenched by strong feedback from the central active nucleus, the latter likely being the

one responsible for the detected FIR emission. We find one possible match among GOODS-N candidates (ID-35028), having a 3.5σ flux at 100 μm and a 1.5σ flux at 160 μm. However, its flux may be strongly blended with that of a very close-by source, only 0.8 arcsec apart, with a lower redshift (z∼ 1.8, more common for Herschel detections) and a low level of on-going SF (0.2 Myr−1according to the fit with standard exponentially declining star formation histories, or τ -models).

For the other fields we used the HerMES DR4 (COSMOS and EGS) and DR3 (UDS) catalogues (Roseboom et al.2010,2012) for the SPIRE bands, while for PACS bands we used the PEP prior-based catalogues (COSMOS and EGS) and the catalogues compiled by the HerMES team (UDS). The only source possibly showing some FIR emission is ID-2075 in COSMOS: it is associated to a source at a distance of 2 arcsec having a 3σ flux at 250 μm, but the flux estimate could be contaminated by a brighter, star-forming source 6 arcsec apart (according to the fit with τ -models; we remind that the PSF at 250 μm is 18 arcsec).

Given these results, and considering the high fraction of potential mis-associations due to the large Herschel PSFs, and the contami-nation from nearby sources, we can conclude that we did not find any clear evidence for FIR emission for any of our ‘reference’ candidates, with only a few moderately uncertain cases.

3.2 Spectroscopic data

Checking the new CANDELS catalogues to be presented in Kodra et al., we find that the following candidates have spectroscopic redshifts (which were used in our analysis): IDs 2490 and 6539 in EGS, IDs 20589 in GOODS-North, IDs 10578 and 16526 in GOODS-South, and ID 8689 in UDS. Among these, only GOODSS-10578 enters our ‘lines’ selection, while all the other are only in the ‘reference’ sample.

We searched the VANDELS (McLure et al.2018; Pentericci et al.

2018) and VUDS (Le F`evre et al.2015) data bases to visually inspect the available spectra for GOODS-South, finding data for additional sources with respect to the CANDELS catalogue. All the

spec-z’s are consistent with the CANDELS estimates, unless explicitly specified. We already discussed IDs 4503, 9209, and 10578 in M18 (Section 4.4). To these, we can now add from the DR2: IDs 4949 and 5934, which show no evident features; ID-12178, which has a strong emission line and is classified at zspec = 0.56 (while zCANDELS=

3.29), but might be spectroscopically contaminated by a very close companion (see Appendix C); and IDs 16526 and 19505, both showing moderate Lyman-α emission, but no other evident features.

(6)

Figure 1. Positions of the passive candidates in the five CANDELS fields (shown is the detection band, WFC3 H160). The dimensions of the images are not

scaled to the real dimensions of the fields. We checked that the sources close to borders have complete image covering. Snapshots of all the candidate galaxies are given in the on-line supplementary material.

Many candidates in GOODS-South were also observed with MUSE, as part of the MUSE-Wide (Urrutia et al.2019) and MUSE Deep (Bacon et al.2017; Inami et al.2017) GTO programs (while all our COSMOS candidates all fall outside the MUSE-Wide COSMOS footprint). Most of the candidates covered by the 1 h deep extended MUSE-Wide data (GOODS-South IDs: 3912, 4587, 4949, 5934, 6407, 7526, 7688, 8242, 8785, 9209, 12178, 17749, 18180, 19301, 19446) show no signs of Ly α emission, in support of these sources candidacy as passive galaxies. Although in general the spectra are not deep enough to provide conclusive evidence, the absence of detectable emission lines strengthen our conclusions. For IDs 9091, 12178 (already mentioned above) and 13394 there is line emission detected within 1.0, 0.5, and 0.5 arcsec in the MUSE-Wide data, respectively. These emission lines are however associated with different foreground objects at redshifts 1.33, 0.56, and 0.42 (MUSE-Wide DR1 ID: 143003008). Three candidates in GOODS-South fall in the MUSE Deep footprint. ID-15457 has no emission detected in the MUSE Deep data; approximately 0.5 arcsec from ID-16506, a Ly α emitter is detected at z= 3.33. Lastly, ID-10578 is confirmed as an active galactic nucleus (AGN) by the MUSE Deep data, as already discussed in M18.

Based on these results, we conclude that we find no strong spectroscopic evidence to exclude any of our candidates from the selected sample.

4 P R O P E RT I E S O F T H E S E L E C T E D S A M P L E The main physical properties of all the candidate objects, as obtained in the SED-fitting procedure with the top-hat library, are given in Appendix B. We provide snapshots and SEDs of all the candidates in the on-line supplementary material, and we show a few significant examples in Appendix C. Here we give a brief summary of the global properties of the candidates.

First of all, we compared the properties of our 102 ‘reference’ candidates with those from the 3D-HST catalogues (Skelton et al.

2014). Most of the 3D-HST photometric redshifts, obtained with EAZY (Brammer, van Dokkum & Coppi2008), are in reasonably good agreement with the CANDELS ones (z < 0.3), with the exceptions of the 19 sources listed in Table 3 (inconsistencies between the two catalogues can be due to many factors, including the usage of different photometric methods and photo-z codes). None of the discrepant objects has a spectroscopic redshift estimate; most of them have lower photo-z in 3D-HST than in CANDELS, with eight candidates that would be excluded from our selections having

z3D-HST<3 and one (ID-13 in GOODS-North) lacking a reliable

photo-z estimate.

We note that in GOODS-South six out of nine discrepant objects are among our strongest candidates (i.e. they belong to the ‘lines’ selection), with three among them having been individually

(7)

Figure 2. Compared 5σ limiting magnitudes for all the wide bands in the

five fields, as given in the CANDELS papers cited in the text. Highlighted with vertical red strips and larger symbols are the bands that are crucial in our analysis, i.e. from left to right WFC3 H160 (the detection band, which has similar depths in all the fields), Ks, and the first two IRAC channels.

Table 3. CANDELS versus 3D-HST photometric redshift outliers (z

0.3). The symbol† highlights the sources that have z3D-HST<3.

Field IDCAND zCAND ID3D z3D-HST

COSMOS 16676 4.127 19670 3.5446 EGS 25724 3.795 33354 2.8674 GOODSN 13 3.014 5 – GOODSN 1570 3.226 2212 0.1497 GOODSN 5744 3.459 8249 3.0668 GOODSN 10672 6.713 15083 1.7364 GOODSN 13403 3.793 18781 0.5635 GOODSN 21034 3.328 28997 4.2929 GOODSN 28344 4.758 3484 4.3757 GOODSN 35028 3.642 34844 1.2655 GOODSS 2608 3.720 4857 3.4116 GOODSS 3912 3.897 7177 4.9983 GOODSS 7526 3.317 15494 2.647 GOODSS 8785 3.852 17894 2.5811 GOODSS 9209 4.486 18684 4.8638 GOODSS 16506 3.382 30821 0.3737 GOODSS 17749 3.697 32872 3.1606 GOODSS 18180 3.650 33566 3.3366 GOODSS 19301 3.592 35502 4.0423 UDS 25893 4.491 41274 3.8881

confirmed as passive sources by the analysis in S19 on ALMA data: IDs 9209, 17749 and 18180. However, in 3D-HST their

photo-z is still above 3, and all have sSFR < 10−11 yr−1 (estimated with FAST; Kriek et al.2009), therefore being robust candidates in their analysis as well. More in general, while∼ 38 per cent of our candidates having z3D-HST>3 can be labelled as passive using

3D-HST FAST estimates and the hard threshold sSFR < 10−11yr−1, in∼ 40 per cent of cases the sSFR is above 10−10yr−1. We note that in contrast, using CANDELS estimates on SFR and masses, we find∼ 14 per cent and ∼ 47 per cent, respectively.

4.1 UVJ diagram

Fig.3shows the position of our passive candidates on the UVJ diagram. The region of passively evolving objects is delimited as in Whitaker et al. (2011). The grey small dots correspond to the whole sample of z > 3 galaxies in the CANDELS fields, while the passive

Figure 3. UVJ selection diagram. The solid lines enclose the region of

passive sources, as identified in Whitaker et al. (2011). The grey dots: all CANDELS z > 3 sources. The large coloured dots: passive ‘reference’ selection; the dots with a square are also part of the ‘lines’ selection. The rest-frame colours are obtained with τ -models fits. The points follow regular patterns because they correspond to the colours of the templates in the SED-fitting library, which have well determined values.

candidates in the ‘reference’ sample are plotted as large dots; the empty squares mark the candidates belonging to the ‘lines’ selection as well.

We point out that for this plot we use the rest-frame colours obtained fitting their observed SEDs with a standard exponentially declining SFH (τ -models): this choice is motivated by the fact that we want to check whether the UVJ colour selection method, straightforwardly applied, can be considered a reliable approach. However, in M18 we showed that the colours obtained with our top-hat SFHs are quite similar, with small shifts in the V−J colour which we ascribe to the less constrained photometry in the observed redder part of the spectrum (the two 5.6/8.0 μm IRAC bands have the poorest SNR), with respect to the visible and near-infrared (NIR) bands which straddle the rest-frame U−V break at z ∼ 3. We thus showed how the UVJ selection criteria is certainly powerful, but at these high redshifts it can miss a number of interesting candidates, in particular many recently quenched objects which still show bluer colours than the typical red passive galaxies. On the same note, Schreiber et al. (2018b) claimed that the UVJ selection tends to be pure (although with a∼ 20 per cent failure rate) but incomplete, in that a fraction of∼ 40 per cent more quiescent galaxies can be identified using e.g. their specific star-formation rate (sSFR) fitted value.

Here too we see that many candidates lie outside the passive region of the diagram. It is interesting to note that most of the outliers belong to two fields, GOODS-S and GOODS-N, which have high-quality data in the infrared bands, in particular considering IRAC CH3 and CH4: this seems reasonable, since their SED fitting is better constrained, yielding less possible solutions and therefore excluding potential star-forming fits. This seems to indicate that in the other fields we are probably underestimating the real number of passive sources.

More generally, we also note that many objects fall inside the passive region of the diagram, but do not belong to our samples. We ascribe this fact to our stringent criteria, which are tailored to select strictly passive objects, rather than quiescent ones (we emphasize that we use the term ‘quiescent’ to dub sources that retain

(8)

Figure 4. Stellar mass versus redshift of the selected candidates in all

CANDELS fields. The grey points show the full CANDELS catalogues, with ‘delayed’ τ SFH mass estimates from Santini et al. (2015), corrected to account for the Chabrier IMF (our fit assumes a Salpeter IMF). The full circles correspond to the objects from ‘reference’ sample; the empty squares mark the objects from the ‘lines’ selections. For the passive candidates we plot the masses from the best top-hat fit in the ‘reference’ run.

a weak star-formation activity, as opposed to completely ‘passive’ ones). As already noted, it might be too conservative excluding potentially reliable candidates because of the mere existence of a few possible, albeit improbable, star-forming solutions in the SED-fitting procedure.

4.2 Diagnostic diagrams

Fig.4shows the stellar mass versus redshift plane. In this case, for the CANDELS full catalogue we consider the masses obtained in the best-fitting with the ‘delayed’ τ -models, for which SFR∝ (t2/τ )×

exp(− t/τ) (Santini et al.2015), while for the red and dead sample we use the ones of the best top-hat fit from this work. We applied a correction factor of 1.75 to make the CANDELS mass estimates, which assume a Chabrier (2003) IMF, consistent with ours, which assume a Salpeter IMF. Most of the ‘reference’ candidates have

M≥ 1010M

, which is unsurprising since our selection criteria

only select luminous (i.e. high SNR) sources. In particular, we find two candidates at z > 4 with masses above 1011M

: UDS-10430

and UDS-25893.

We then consider the half-light radii of 3 < z < 5 sources, as measured by the SEXTRACTORdetection runs in the H160 band, converting them from pixel space to arcseconds and then to proper kpc at the relevant redshifts by means of appropriate Astropy (Astropy Collaboration2013; Price-Whelan et al.2018) routines.

We do not take into account the k-correction factor which would be necessary since we consider galaxies at different redshifts, implying that the observed H band samples different regions of the rest-frame spectrum (i.e. 400 nm at z= 3, 320 nm at z = 4 and 267 nm at z= 5). However, we have checked that the sizes measured on images corresponding to bands sampling the same region of the spectrum (i.e. Y-band at z= 3 and J-band at z = 4, at 267 nm rest-frame; all images were PSF matched to H) are very similar to the H band sizes, and considering that the Y-band data are not always available we preferred to simply use H band data. A possible drawback of this choice is that in this way we are looking

Figure 5. Top: mass–radius relation. The grey dots: all CANDELS sources

at 3 < z < 5; the coloured circles: passive candidates, size-coded for redshift (see legend); empty squares mark the objects in the ‘lines’ selection. The radii are the half-light radius in H band from theSEXTRACTORruns, converted into proper kpc. The masses are the ‘delayed’ τ SFH estimates from Santini et al. (2015), corrected to account for the Chabrier IMF (our fit assumes a Salpeter IMF). Most of the candidates are on the compact tail of the distribution, with a clear trend with redshift: high redshift sources are smaller compared to z∼ 3 sources, as shown in the bottom panel, where we plot the radii of the candidates against their redshift.

at the rest-frame B band, which is not particularly well suited to study the size of a red object.

The resulting mass versus radius relation is plotted in Fig.5. In general, the passive candidates appear to be typically massive but mostly concentrating towards the region of small radii. Also, a dependence on the redshift seems to be present, as highlighted by the size-coding of the dots and in the lower panel of the figure: earlier sources appear to be more compact than later ones. The typical radii of these objects are roughly consistent with the ones found for other high-redshift samples of passive objects, Re∼ 1 kpc, therefore

showing a general compactness with respect to local galaxies (e.g. Cimatti et al.2008, for 1.4 < z < 2 passive galaxies in a similar mass range).

4.3 A massive red and dead galaxy at z∼ 6.7?

We find a particularly interesting object among the galaxies in our selection: GOODS-North ID-10672 has zCANDELS= 6.713, and it

enters both the ‘reference’ and ‘lines’ samples. Its mass inferred from the SED-fitting is 4× 1010M

, with an age of∼500 Myr,

implying a formation redshift of∼14, i.e. ∼300 Myr after the big bang; the fitted burst duration is of just 300 Myr, implying an average SFR of 130 Myr−1. With a typical τ -model SFH, it is fitted with sSFR= 5.21877 × 10−11yr−1, so it would be classified as quiescent, but it would fail a strictly passive selection based on a standard sSFR

<10−11yr−1criterion. The object has been observed and catalogued in many surveys (e.g. Bouwens et al.2015; Finkelstein et al.2015;

(9)

Table 4. Matches between the ZFOURGE/3D-HST and CANDELS IDs in

the Schreiber et al. (2018b) sample.

ZFOURGE / 3D-HST CANDELS ZF-COS-20115 COS 16676 3D-EGS-18996 EGS 14727 3D-EGS-31322 EGS 24177 ZF-UDS-3651 UDS 1244 ZF-UDS-4347 UDS 2571 ZF-UDS-7329 UDS 7520 ZF-UDS-7542 UDS 7779 ZF-UDS-8197 UDS 8682 ZF-UDS-41232 UDS 25688

Harikane et al.2016) as a z∼ 6−7 mildly star forming or quiescent galaxy. As already pointed out (Table3), in 3D-HST it is instead classified a z= 1.73 source by means of EAZY photometric redshift estimate (without grism information). We find that no detectable signal is associated to this galaxy in MIPS, PACS, and SPIRE (i.e. 24 to 500 μm): while some flux is detected within a few arcsec from the source position, it can quite confidently be attributed to other nearby star-forming objects. However, from a SED-fit with redshift fixed at the 3D-HST value of 1.73 we obtain SFR values of ∼ 3 − 8 Myr−1(depending on the details of the adopted models),

which is too low to be detected in MIPS or Herschel. Another possibility might be that the source is actually a cold star, possibly a brown dwarf; instead, we tend to rule out the options of a young and obscured star or of an AGB star, because the source is very isolated, which would be unusual for these kind of objects. A spectroscopic observation would be important to definitively rule out alternative possibilities, but the very faint continuous emission would make the analysis very difficult and perhaps unfruitful. Snapshots and best-fitting SED of this objects are shown in Appendix C; a more extended analysis to exclude the possibility of degeneracies in the fitting would be required, but is beyond the scope of this work and we leave it to future work.

4.4 Comparison with other selections

In M18 we compared our results on GOODS-South to the ones by Straatman et al. (2014, S14) and Nayyeri et al. (2014, N14). We repeat the analysis now, given that our sample has sligthly changed because of the new redshift estimates. Cross-correlating the CANDELS catalogue with the ZFOURGE selection in S14, we now find that four out of six sources in the S14 selection belong to our sample as well: ID-19883 has now entered both our ‘reference’ and ‘lines’ selections, together with IDs 4503, 17749, and 18180 which were already present in M18. As for the N14 selection, the situation in unchanged, with five sources in common and the other 11 rejected in our analysis. We refer the reader to M18 for other considerations about these comparisons.

More recently, Schreiber et al. (2018b) used MOSFIRE H and

K spectra to confirm the quiescent nature of 22 galaxies (out of

24 reduced spectra) detected in COSMOS, EGS, and UDS, based on ZFOURGE and 3D-HST catalogues. These sources had been previously singled-out as quiescent candidates via UVJ selection. We cross-matched the ZFOURGE, 3D-HST, and CANDELS cata-logues RA-DEC coordinates, and found that out of the 22 objects, only nine are in our selection as well (see Table4). We could not find matching objects for three spectroscopic sources (ZF-COS-17779 is outside the HST footprint, ZF-UDS-35168 and ZF-UDS-39102 are not detected in CANDELS); among the other nine, two have

zCANDELS <3, while the other seven do not match our selection

criteria (having some SF solutions; we must point out that they select quiescent rather than totally passive sources).

5 N U M B E R D E N S I T Y

We now try to estimate the number density of passive galaxies, dividing the number N of our red and dead candidates observed within a redshift interval of interest z by the cosmological volume

V corresponding to such interval, re-scaled by the ratio between the

total area of the survey (∼969.7 sq. arcmin) and that of the full sky. We compute the errors as the Poissonian uncertainties√N /V, which we assume do take into account cosmic variance given the fact that we are considering five different realizations (fields). As a reference result, we first compute the number densities yielded by our list of candidates, without considering any correction for incompleteness; we will discuss and complement this in the next subsection.

The resulting number density of passive galaxies in the redshift interval 3 < z ≤ 5, considering the sum of our samples in the five fields, is 1.73 ± 0.17 × 10−5 for the ‘reference’ selection (6.69± 1.08 × 10−6for the ‘lines’ selection) Mpc−3. These values would change to 2.16± 0.22 × 10−5(8.52± 1.54 × 10−6) Mpc−3 if we excluded COSMOS from the average, a potentially reasonable choice given that this field yields significantly lower estimates with respect to the other four (which is probably due to cosmic variance, but also to worse photometric properties like e.g. the broad FWHM in Ks, see Section 2.1). However, in the following we will stick to the estimates obtained averaging on all the five fields.

5.1 Completeness

It is not easy to try and quantify the completeness of our sample of passive candidates, given the different depths of the five fields, and the subtleties of the technique we adopted. We attempt to do so proceeding as follows. UsingZPHOT, we create a library of 3000 3 < z < 5 synthetic spectra with top-hat SFHs, having different ages and duration of SF bursts so that 1740 are passive and 1260 are star forming. Then, we consider the two fields with the highest and the lowest quality of data in terms of depth, namely GOODS-South and EGS, respectively; by means of the in-house software

SIMULCAT, we use the synthetic models to create two simulated observed catalogues, having the properties of the two fields in terms of filters and depths (i.e. signal-to-noise ratios) at all magnitudes and in all the observed bands, and fixing a reference filter (we choose IRAC-CH2) to predefined magnitude values, i.e. mI2= 21, 22, 23,

24, 25, 26, 27. In both cases, each model is replicated ten times with a slightly different noise realization, so each catalogue finally contains 17400 passive objects. Finally, we fit these mock observed catalogues with our top-hat library, and proceed to select passive candidate following exactly the same procedure we used on the real data.

The results for the two fields are show in Fig.6. In both panels, the solid lines show the inferred completeness in four redshift bins, as a function of the reference magnitude in IRAC-CH2. This completeness is the result of the product of two factors: the photometric completeness, which comes from the pre-selection we perform on the observed catalogue, as described in the second bullet of the list at the beginning of Section 2.1 (namely: H < 27,

SNRK, IRAC1, IRAC2>1, plus an additional condition on the SNR of

the detection band to take into account the typical detection cut in the original input catalogues, SNRH>4), shown as dashed lines in the

(10)

Figure 6. Estimated completeness of the passive selection as a function

of IRAC-CH2 observed magnitude, as estimated from a set of dedicated simulations, for the two fields with the best and worst data quality (upper panel: GOODS-South; lower panel: EGS). In each panel, the dashed lines correspond to the photometric selection completeness, the dotted lines to the passive candidates selection completeness (computed considering only the photometrically selected galaxies), and the solid lines to the total completeness given by the product of the two. The colours of the lines correspond to different redshift bins. See the text for more details.

Figure; and the passive-selection completeness, which we computed considering only the sources that survived the photometric selection, and is shown with dotted lines. Therefore, the product of the two quantities is self-consistent as a total completeness.

As expected, the values drop as a function of the magnitude and of the redshift. In the case of GOODS-South, a 50 per cent completeness is reached at mIRAC2  24.75 at 3 < z < 3.5, and

at mIRAC2  23.75 at 4.5 < z < 5; in EGS, the 50 per cent

completeness is reached at∼1 magnitude brighter. We note that we are∼ 95 per cent complete in GOODS-South at mIRAC2 23 at

z= 3.

We now use these values to estimate an inferred ‘true’ number densities of sources (we only do so for the ‘reference’ selection, for the sake of simplicity). We multiply the number of objects we actually find in each magnitude and redshift bin by the inverse of the corresponding completeness estimate; as a reasonable approx-imation of the complex features of the different fields, we use the values that we obtained using GOODS-South for the two GOODS catalogues, and the estimates computed on EGS for the other three. To obtain a fair estimate, we take 50 per cent as the minimum reliable value, and when the completeness drops below this threshold we continue to use it to compute the actual multiplicative factor. Also, if a bin of magnitude and redshift contains zero objects, its counts will of course remain zero. These two points imply that once again we are being conservative, and the obtained number density can still

Table 5. Number densities (Mpc−3) of passive candidates in the total

CANDELS area (∼969.7 arcsec2), considering three redshift intervals, for our two selections ‘reference’ (both as observed and after correction for completeness) and ‘lines’.

z Reference Lines

Observed Corrected

3 < z≤ 5 1.73× 10−5 2.30× 10−5 6.69× 10−6 3 < z≤ 4 2.90× 10−5 3.66× 10−5 1.08× 10−5 4 < z≤ 5 4.34× 10−6 7.94× 10−6 2.17× 10−6

Table 6. Number densities (Mpc−3) of passive candidates in the five

CANDELS fields, considering the cosmological volume 3 < z ≤ 5, for our two selections ‘reference’ (both as observed and after correction for completeness) and ‘lines’.

Field Reference Lines

Observed Corrected COSMOS 3.08× 10−6 3.50× 10−6 1.54× 10−6 EGS 1.05× 10−5 1.34× 10−5 4.04× 10−6 GOODS-N 3.36× 10−5 4.38× 10−5 9.62× 10−6 GOODS-S 3.17× 10−5 4.44× 10−5 1.25× 10−5 UDS 1.32× 10−5 1.75× 10−5 7.42× 10−6

be underestimated. With this approach, we end up with a corrected total number density of 2.30± 0.20 × 10−5for the whole redshift interval 3 < z < 5 (for the sake of reference, we point out that if we decided not use the 50 per cent completeness threshold, and instead we used the full completeness functions as obtained from the simulations, we would get a number density of 3.25± 0.24 × 10−5). In Table5we list the number densities both before and after the completeness correction, for the whole survey area, in three redshift intervals; while in Table6we give the number densities corresponding to the individual cosmological volumes of the five fields, in the full redshift interval 3 < z≤ 5. GOODS-South yields the higher value (by a factor of∼2 with respect to the average of the five fields, in the ‘reference’ selection), likely because of a better constrained photometry in the infrared bands leading to more robust SED fits. Cosmic variance also plays a role: the number of all z > 3 sources varies by a factor of∼1.5 between the five fields (see Table2).

5.2 Comparison with predictions from numerical simulations To understand how the values we have obtained fit in the current theoretical scenario, we have compared the number density of our passive candidates at high redshift with the estimates obtained in five recent cosmological hydrodynamical simulations: ILLUSTRIS

(Vogelsberger et al.2014b,a; Genel et al.2014; Nelson et al.2015),

ILLUSTRIS-TNG100 and TNG300 (Pillepich et al.2018; Nelson et al.

2019),EAGLE(Schaye et al.2015), andSIMBA(Dav´e et al.2019). Full simulations data containing particles and groups information are publicly available and downloadable for the first four, while

SIMBAdata were privately provided by the authors.

The ILLUSTRIS and ILLUSTRIS-TNG simulations exploit the moving-mesh codeAREPO(Springel2010); simulated volumes and baryonic mass resolutions are as follows: 106.53Mpc3, 1.26×

106M

 forILLUSTRIS; 110.73Mpc3, 1.4× 106M

forILLUSTRIS -TNG100; 302.63Mpc3, 1.1× 107M

forILLUSTRIS-TNG300. The

EAGLEsimulations exploitGADGET-3, the latest incarnation of the original Tree-SPH (smoothed particle hydrodynamics) code

(11)

oped by Springel (2005); we used data from the RefL0100N1504 simulation, which is the most complete in the suite of runs in terms of physics included in the code; the run simulates a volume of 100.03Mpc3, and has a baryonic mass resolution of 1.81× 106M

.

Finally,SIMBAis based on theGIZMOcode by Hopkins (2015), a mesh-free finite-mass hydrodynamic code which handles shocks via Riemann solvers, with no need for artificial viscosity; the simulation also includes detailed and novel recipes for AGN feedback, and has a volume of 1003Mpc3, with a baryonic mass initial resolution

of 1.82× 107M

, i.e. comparable to the TNG-300 simulation. All the models include baryonic sub-grid physics to simulate star and black hole formation, stellar and AGN feedback, and metal enrichment; whileILLUSTRIS, ILLUSTRIS-TNG, and SIMBAmodel sub-grid physics from first principles,EAGLEuse empirical relations to match observed properties of galaxies at z= 0. All simulations assume a Chabrier IMF, and we applied again the corrective factor of 1.75 to compare with the Salpeter IMF adopted in our SED-fitting procedure. We also paid attention to consider the cosmological factor h= H0/100 in the computation of masses and SFRs, when

necessary (some simulations define masses and lengths in units of

h, others do not).

For the models, we considered simulated galaxies with

M/M>5× 109, and we adopted the usual selection criterion

sSFR < 10−11yr−1. Since in theILLUSTRIS andILLUSTRIS-TNG data releases multiple values of masses and SFRs are given for each simulated object, depending on the different radii within which they are computed, we tried to mimic as accurately as possible the observational approach: to this aim, we used the values estimated within twice the half-mass radius of each object (see also the analysis in Donnari et al.2019). On the other hand,EAGLEoutputs masses and SFRs within a set of apertures (defined as the diameters of spheres centred on the position of the object) of fixed proper lengths, from 1 to 100 proper kpc; and two estimations of the half-mass radius, R1/2,30and R1/2,100, computed within 30 and 100 kpc,

respectively. To obtain a fair comparison with theILLUSTRISand TNG cases, in this case we compute the sSFR by considering, for each simulated galaxy in a snapshot, the mass and SFR within the aperture which is closest to 4× R1/2,100. Finally, the values of the

sSFR within twice the half-mass radius of each object forSIMBA

were directly communicated by the authors.

In Fig.7we summarize the results of our analysis on number densities. We show three estimates from other studies on observed data (empty squares): Muzzin et al. (2013) use the colour–colour

UVJ diagram to select quiescent galaxies in COSMOS/UltraVISTA,

in seven redshift bins up to z∼ 4 (we show the values corresponding to the M/M>1010selection, in grey); Straatman et al. (2014)

select quiescent sources in the ZFOURGE survey, again using the

UVJ criterion, and focusing on the mass range log(M/ M) > 10.6 and the redshift range 3.47 < z < 4.25 (in red – but we report their lower redshifts estimates as well); and Schreiber et al. (2018b) estimate the number density of quiescent galaxies at 3 < z < 4 and M/M>3× 1010 by combining the UVJ selection on

ZFOURGE photometric data with MOSFIRE spectral analysis, and we plot their two estimates for strictly colour-selected objects (large empty cyan square) and sSFR-threshold candidates, which include recently quenched galaxies, as in M18 (smaller cyan square). We then plot the results based on our SED-fitting method for the ‘reference’ (both before and after the correction for incompleteness, as circles or diamonds, respectively) and the ‘lines’ selections (as stars), considering both the cuts in mass and redshift adopted by S14 (red empty symbols), and the 3 < z < 5 interval without any mass cut (dividing in two redshift bins, 3 < z < 4 and 4 < z <

Figure 7. Number density of passive galaxies as a function of redshift.

We plot many different estimates, both from observations (Muzzin et al.

2013; Straatman et al.2014; Schreiber et al.2018b) and from hydrody-namical simulations (from which we take the galaxies with M/M>

5× 109):ILLUSTRIS-1 (Vogelsberger et al.2014a),ILLUSTRIS-TNG100, and ILLUSTRIS-TNG300 (Pillepich et al.2018),EAGLE(Schaye et al.2015, run RefL0100N1504 with full physics included), andSIMBA(Dav´e et al.2019). See the text for relevant information on these data sets and models. We then plot the results based on our SED-fitting method for the ‘reference’ sample, considering the cuts in mass and redshift adopted by S14 (red empty circle), and the full 3 < z < 5 interval divided in two bins considering all masses, before (black circles) or after (black diamonds) correcting for completeness as described in Section 5.1. We also plot the values corresponding to the ‘lines’ selections (black or red stars). The shadowed area highlights the relevant redshift interval. The agreement between all observations and the

ILLUSTRIS-TNG models is reasonable up to z  4, while it worsens at earlier epochs; on the other hand,EAGLEyields better results at z > 4, but underestimates the number of passive objects at z < 3. Finally,SIMBAseem to fall short at all epochs, however doing better than the originalILLUSTRIS-1 model. See the text for more details.

5 – black solid symbols). Finally, we plot the results from the five hydrodynamical simulations (solid coloured lines), as described above, showing also a case in which we consider the total sSFR of the simulated galaxies to select the passive ones, rather than the sSFR of the central regions, for the sake of comparison.

Looking at the resulting plot, at 3 < z≤ 4 we find a reasonable agreement between our number density estimations, the ones obtained by other observational studies, and theEAGLEand TNG-100 models. TNG-300 is close enough as well, although its number density is slightly lower, perhaps because of the lower resolution of the simulation with respect to TNG100. Remarkably, while the originalILLUSTRISsimulation is not able to reproduce the properties of the galactic populations at high redshift and falls short at all epochs above z > 1 in reproducing the number of quenched galaxies, the TNG runs have largely cured this issue. On a side note, we point out that if we only considered the CANDELS fields yielding the highest densities the agreement both with other observations and with the simulations would be much less satisfying. We also see that, consistently with what we found in M18, our estimate is lower

(12)

than the one by S14 using their cuts for mass and redshift, and we ascribe this to their shallower selection criteria, which also include mildly star-forming objects in the selection.

On the other hand, at z > 4 the TNG models still show a clear tension with the observational data, whereas theEAGLEmodel performs better; however, theEAGLEtrend is too flat and struggles at reproducing the observed number densities at z < 3. It is difficult to trace back the origin of this different trends to the various physical mechanisms in the models, because hydrodynamical simulations are highly non-linear by construction. It is worth pointing out that the ILLUSTRIS and TNG mass functions are typically close to the observed ones (e.g. Genel et al. 2014; Pillepich et al.

2018), pointing to the conclusion that the discrepancy at z > 4 is not due to a poor sampling of the whole high-redshift galaxy populations, but more likely to enduring issues with the modelling of quenching mechanism. Gas hydrodynamics, radiative cooling, star formation and feedback from stars and AGNs are indeed implemented in different ways in TNG andEAGLE. We speculate that details of the AGN feedback implementation play a major role in regulating the activity of the simulated galaxies, and therefore in defining the properties of the passive populations at different redshifts. In particular, the thermal feedback implemented inEAGLE

is more efficient at high redshift, while in TNG the most effective mechanism is the kinetic feedback, which has a larger impact at low redshifts (Weinberger et al.2018), consistently with the global trends we have found. However, these are basic speculations. A detailed work specifically dedicated to the analysis of this topic is currently in the making (Fortuni et al., in preparation).

We point out that the values of SFR and masses that we choose to consider play a major role in reconciling models and observations: indeed, for TNG and EAGLE the sSFRs computed over limited central areas turn out to be typically lower than those computed considering the full extension of the objects, and in some cases they are exactly zero, thus allowing the inclusion of more sources in the passive selection. On the other hand, using the simplest approach of summing on all the particles of a simulated galaxy would lead to strong tensions at all redshifts, as in many cases the sSFRs would be higher, excluding objects from the selection criteria and finally resulting in lower number density estimates (see e.g. the blue dashed line in Fig.7, which shows the case for TNG100 model). Note that this implies that in the models the stellar mass profiles are more centrally concentrated than SFR profiles (at least at high-z), so that a significant amount of star formation takes place in the outskirts of the galaxies (see also Donnari et al.2019): therefore, enlarging the radius in which the SFR and mass are computed yields higher sSFR values that eventually exclude many objects from the passive selection. If this is the case, current observations might underestimate the actual cosmic SFR, missing some amount of peripheral activity.

Finally, the number densities inSIMBAare closer to the observed values than the ones fromILLUSTRIS-1, but fall short with respect to TNG and EAGLE. Remarkably, in this model the values remain almost unchanged varying the aperture radius over which SFRs and masses are computed, at variance with the other simulations. Apparently, the peripheral star formation activity is not present in

SIMBA. This might be due to the different star-formation prescrip-tions in the codes: whileILLUSTRISand TNG assume that stars can form at gas densities n > 0.13 cm−3(Vogelsberger et al.2013), in

SIMBAa sub-grid criteria based on the estimated local density of the H2 molecule is adopted (Dav´e et al.2019), and this tends to require significantly global higher densities to trigger the activity (Dav´e, private communication). As a result, TNG might form more

stars in the outskirts of galaxies thanSIMBA, perhaps in particular as a result of shock-induced cooling and collapsing of gas. On the contrary, the resolution of the simulations seem to play a minor role: while the TNG-100 andEAGLEones are finer (∼ 106M

 per gas

particle), the TNG-300 one is comparable toSIMBA(∼ 107M per

gas particle), but its number density of passive sources is close to the one in TNG-100.

As final remarks, it must be pointed out that: (i) we are applying demanding criteria which only include in the sample very robust candidates; (ii) since we are working with the CANDELS cata-logues, we are not considering H-dropouts; K or IRAC-detected sources might increase the actual number of high-redshift passive galaxies, since sources that quenched long before the observa-tion have negligible UV rest-frame flux, thus failing the H-band detection at high redshifts; (iii) as shown in Section 5.1, the observed values are most likely far from completeness, and although we also considered ‘corrected’ values, we might anyway still be underestimating the total number of real passive sources. These points suggest that the tension with models cannot be considered completely ruled out.

We point out that the volumes of the simulations are comparable to the two considered cosmological volumes for 3 < z < 4 and 4

< z <5, which are of the order of 3× 106Mpc3; in particular,

while TNG-100,EAGLEandSIMBAare a factor of∼3 smaller than the observed volumes, TNG-300 is a factor of∼9 larger, and is therefore statistically significant.

6 T H E C O N T R I B U T I O N O F T H E PA S S I V E P O P U L AT I O N T O T H E U N I V E R S A L S TA R F O R M AT I O N R AT E D E N S I T Y

The SED-fitting approach that we adopted in this study has another advantageous outcome. The best-fitting template of each galaxy can be used to infer the complete description of its SFH through the cosmic history, by means of two of its free parameters: τ (the

e-folding time for the exponentially declining models, or the burst

duration for the top-hat models) and C (the normalization factor of the solution). They are the only two parameters needed to constrain the functional form of the SFH, (z). From it, we can compute the average SFR of a galaxy in any redshift interval, and then obtain an estimate of the universal SFRD summing up all the contributions from individual galaxies.2

As a cautional remark, it is worth mentioning that this approach implies the assumption that the best template in the SED-fitting process (the one with the lowest χ2) is the ‘correct’ one, while in

reality many more or less plausible solutions can be associated to each source. We do not take into account the uncertainties associated to this complication. In addition to this, we should also recall that it is assumed that galaxies evolve in isolation; in other words, mergers

2Starting from the fitted quantities, the procedure is as follows. Each source’s

best-fitting model is characterized by the stellar mass Mat the age the galaxy has when it is observed, tf. From these two fitted quantities, one can infer

the normalization factor C as

C=tf M∗ 0 f(t)× (t)dt

, (1)

where f(t)= 1 − frec(t) is the dimensionless parameter that accounts for

the recycled fraction of gas, and (t) is the instantaneous normalized SFR (given by the SFH model, for example (t)= exp(− t/τ) for the exponentially declining models). The true instantaneous SFR at any epoch

t is then (t)= C × (t).

(13)

Figure 8. Sketch of the method adopted to compute the global SFRD across the cosmic time. We sum the contributions of all the galaxies observed in a given

redshift interval dzobsto the SFR in each redshift bin dzHacross the cosmic history. Such contributions can be evaluated from the shape and normalization of

the SFH of each galaxy, which in turn come from the SED-fitting procedure. Then, we divide the obtained total SFR in each bin dzHby the dzobscomoving

volume to obtain the SFRD in that dzobs. Finally, we average over the dzobsbins to obtain the SFRD of each dzH. are not considered in this approach. This is by construction inherent

to the SED-fitting method. However, we do not consider this point to be invalidating: the past evolution of an object and its SFH can be caused by a number of causes/events, but the amount of stars formed per year is the same, independently from the merger history.

To assess the reliability of the method, we first compare the reconstruction of the SFRD based on the SED-fitted SFHs of the CANDELS catalogues with the observed ones; we consider the well-known one by Madau & Dickinson (2014, M14), and the one by Yu & Wang (2016, Y16, obtained from a Markov chain Monte Carlo (MCMC) technique applied to the observed mass functions rather than from the direct summation of UV and IR fluxes at the observed redshifts). For the CANDELS data, we use the delayed

τ-models (in which the star formation is parametrized as SFR(t)=

t2 × exp(− t/τ)) and the mass estimates with method 6a delτ,

presented in Santini et al. (2015).

We point out that we cannot apply any correction for incom-pleteness here, since the approach is based on the computation of individual SFHs, which by definition cannot be inferred for undetected objects. In this exercise, the analysis is necessarily limited to galaxies with H160 < 27.5, i.e. to a certain mass cut, while the SFRD is typically computed by integrating the observed luminosity and mass function down to very faint limits in SFR and mass, far below the observed one, by deep extrapolation. Therefore, our estimation is actually a lower limit.

We proceed as follows. First, we define two grids of redshift bins: (i) dzobs, which we use to bin the observed catalogue. Each

bin can represent a slice of Universe which evolves comovingly, and therefore its global SFH can be considered an independent realization of an universal SFH, for z > zobs;

(ii) dzH, which we use to bin the cosmic history: we will obtain

the global SFH of each dzobs, as the sum of the SFRs of individual

galaxies (computed as described above) in each dzH.

In other words, we select from the CANDELS catalogues the galaxies observed in a bin dzobs, and we trace back their SFH,

summing all their contributions in each dzH bins. To obtain the

SFR density, we then divide the result by the comoving volume of dzobs: i.e. the volume of the spherical shell centred on the observer,

extending from zobsto zobs+ dzobs, and re-scaled to the total angular

area of the five fields. We then repeat the procedure for all the dzobs;

finally, in each dzHwe compute the mean of the various SFRDs

corresponding to each dzobs. In formulae:

∀dzobs−→ SFRdzH =

n



i=1

SFRdzH,i, (2)

where n is the number of galaxies in the considered dzobs; then,

∀dzobs−→ SFRDdzH= SFRdzH/Vdzobs (3) and finally SFRDdzH = Ndzobs j=1 SFRDdzH,j Ndzobs . (4)

The method is sketched in Fig.8. In Fig.9we plot the result of this approach, together with the two cited observational ones.

Our points are in good agreement with the curves, and particularly with Y16, although they are consistent with both estimations con-sidering the uncertainties (the error bars are the standard deviations of the distribution of the SFRD from each dzobs). We speculate that

while the discrepancy with the M14 curve can be understood in terms of the caveats on the incompleteness of our sample discussed above, the good consistency at z < 5 with the Y16 estimate is justified by the fact that we are basically adopting their same method to compute the SFRD (see their equation 5). Therefore, the SFR estimated from the observed UV/IR fluxes seems not strictly equivalent to the SFR inferred from the observed stellar mass, which might be due to several factors discussed in Y16, among which the IMF, the metallicity, the outshining of young stars, or overestimated absorption, as well as possible differences in the computation of the recycled fraction. However, the tension with MD14 is not dramatic, and furthermore, we note that the precise values of the averaged SFRDs depend on the choice of dzobsbinning

(and binning is sinning...).

At z > 5 our estimation apparently diverges from the other two, but the scatter between the SFRDs from the various dzobsbins is

large enough to make the global estimate consistent with the curve within the error budget. We recall that the bars correspond to the standard deviation of the distribution of the individual SFRDs of each dzobs, so that their extension at high z is likely due to poor

statistics (few detected objects, plus cosmic variance); on the other

Riferimenti

Documenti correlati

Since the latter equation is nonlinear and we wish to solve it by means of a perturbation argument which allows us to obtain a new solution which is as close as we want to the

On behalf of the International Observatory on Participatory Democracy (IOPD), we congratulate you, especially this year which had 32 nominations of high level.. The award will

La rivalutazione dell’importanza dei sensi e della loro intera- zione sinestetica (Riccò, 1999) nella nostra esperienza conoscitiva è tutto sommato recente e corregge gli interessi

An Australian study noted over a decade ago that evaluations in the 1990s did not establish if that country’s science awareness programme “caused Australians to become more or

While sites like Airbnb and Uber became giant companies, the platform on which we would share our power drills with neighbors never took off.. Instead of platforms that would

We will relate the transmission matrix introduced earlier for conductance and shot noise to a new concept: the Green’s function of the system.. The Green’s functions represent the

Tuttavia, il prezzo predatorio rappresenta un dilemma che ha tradizionalmente sollecitato l’attenzione della comunità antitrust: da un lato la storia e la teoria economica